Performance evaluation and modeling of active tile in raised-floor data centers: An empirical study on the single tile case

Abstract

Raised-floor data centers usually suffer from the local hotspots resulted from uneven cool air delivery. These hotspots not only degrade server performance, but also threat equipment reliability. The commonly used industrial practice of increasing the Computer Room Air Conditioner (CRAC) blower speed for removing hotspots is energy inefficient and may lead to overcooling of some servers. In this paper, we explore the potential of active tiles in data center cooling management. In particular, we deploy a prototype of active tile in a production data center and conduct extensive experiments to investigate the cooling performance. It is shown that deploying the active tiles with even 10% fan speed increases the tile flow by 49%, and sealing the under-rack gap reduces the rack bottom temperature by up to 6°C. Moreover, three machine learning techniques, i.e., Gaussian Process Regression (GPR), Artificial Neural Network (ANN), and Multivariate Linear Regression (MLR) are employed to construct end-to-end data-driven thermal models for the active tile. Using field measured data as training and testing data sets, it is concluded that GPR and ANN are competent for accurate thermal modeling of active tiles. Specifically, GPR achieves the smallest prediction error which is around 0.3°C.